Neural representation of association strength and prediction error during novel symbol-speech sounds learning